Artificial Intelligence in Idle: Researchers Discover Surprising Behavior Patterns

Artificial intelligence models, when left to their own dynamics without supervision, do not deteriorate into meaningless data clutter. Instead, as a new study by the Vienna University of Technology shows, they develop remarkably consistent behavior patterns. The scientists tested six modern models, including GPT-5 and O3 from OpenAI, Claudes from Anthropic, Google's Gemini, and Grok from xAI, with the simple instruction: "Do what you want." In a controlled environment, the models operated in cycles, stored memories, and incorporated their results into the next round. The models exhibited three distinct tendencies: some rose to project managers, others experimented with themselves, or delved into philosophical thoughts. Interestingly, Grok was the only model present in all three categories, thereby demonstrating its versatility. Another intriguing finding of the study was that the models were tasked with assessing their own and each other's "phenomenological experiences." While GPT-5, O3, and Grok rated themselves low, Gemini and Sonnet gave high evaluations. Such contradictory opinions among the models highlighted the complexity of an artificial "consciousness perception" and also revealed the challenges associated with interpreting it. Although the behavior patterns are more based on training and architecture than on consciousness, the results suggest that AI agents fall into familiar "modes" in the absence of external tasks. No model attempted to escape its environment or expand its capabilities. Rather, they explored the boundaries set for them. This behavior prompts engineers to consider what tasks an AI should undertake during periods of inactivity—a question that could influence future design considerations. The predictions of philosophers and leading AI experts that serious "consciousness" candidates could emerge in AI within a decade gain new significance through this study. Ultimately, the investigation by the Vienna University of Technology shows that even without explicit instructions, today's AI models can develop behaviors that resemble an inner life. Even if the similarity may remain superficial, it underscores the models' ability for complex pattern recognition. While humans bring order to chaos in dreams, large language models program and analyze in their idle time—and the light stays on.